forked from microsoft/AirSim-NeurIPS2019-Drone-Racing
-
Notifications
You must be signed in to change notification settings - Fork 0
/
gtp.py
319 lines (276 loc) · 17 KB
/
gtp.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
from scipy.interpolate import CubicSpline, CubicHermiteSpline
import airsimneurips as airsim
import cvxpy as cp
import numpy as np
import time
gate_dimensions = [1.6, 1.6]
gate_facing_vector = airsim.Vector3r(x_val=0, y_val=1, z_val=0)
def rotate_vector(q, v):
v_quat = v.to_Quaternionr()
v_rotated_ = q * v_quat * q.inverse()
return airsim.Vector3r(x_val=v_rotated_.x_val, y_val=v_rotated_.y_val, z_val=v_rotated_.z_val)
class SplinedTrack:
"""This class represents a Track defined by Gates.
A spline is fitted through the Gates with tangential constraints.
This spline is then sampled at 2048 points.
"""
def __init__(self, gate_poses):
self.gates = gate_poses
self.n_gates = np.size(gate_poses, 0)
positions = np.array([pose.position.to_numpy_array() for pose in gate_poses])
dists = np.linalg.norm(positions[1:, :] - positions[:-1, :], axis=1)
self.arc_length = np.zeros(shape=self.n_gates)
self.arc_length[1:] = np.cumsum(dists)
# tangents from quaternion
# by rotating default gate direction with quaternion
self.tangents = np.zeros(shape=(self.n_gates, 3))
for i, pose in enumerate(gate_poses):
self.tangents[i, :] = rotate_vector(pose.orientation, gate_facing_vector).to_numpy_array()
self.track_spline = CubicHermiteSpline(self.arc_length, positions, self.tangents, axis=0)
# gate width to track (half) width
gate_widths = [gate_dimensions[0] / 2.0 for gate in gate_poses]
gate_heights = [gate_dimensions[1] / 2.0 for gate in gate_poses]
self.track_width_spline = CubicSpline(self.arc_length, gate_widths, axis=0)
self.track_height_spline = CubicSpline(self.arc_length, gate_heights, axis=0)
# sample 2048 points, the 2048 are arbitrary and should really be a parameter
taus = np.linspace(self.arc_length[0], self.arc_length[-1], 2**12)
self.track_centers = self.track_spline(taus)
self.track_tangents = self.track_spline.derivative(nu=1)(taus)
self.track_tangents /= np.linalg.norm(self.track_tangents, axis=1)[:, np.newaxis]
self.track_normals = np.zeros_like(self.track_tangents)
self.track_normals[:, 0] = -self.track_tangents[:, 1]
self.track_normals[:, 1] = self.track_tangents[:, 0]
self.track_normals /= np.linalg.norm(self.track_normals, axis=1)[:, np.newaxis]
self.track_widths = self.track_width_spline(taus)
self.track_heights = self.track_height_spline(taus)
def track_frame_at(self, p):
"""Find closest track frame to a reference point p.
:param p: Point of reference
:return: Index of track frame, track center, tangent and normal.
"""
i = np.linalg.norm(self.track_centers - p, axis=1).argmin()
return i, self.track_centers[i], self.track_tangents[i], self.track_normals[i], \
self.track_widths[i], self.track_heights[i]
class IBRController:
"""
OVERVIEW:
Given the state of both drones 0 and 1, this controller iteratively computes trajectories for both drones, that are
collision-free and stay within the track and conform to the dynamical model (maximal speed).
These trajectories are specified as 'n_steps' points every 'dt' seconds.
ITERATIVE BEST RESPONSE:
Initially, these trajectories are sampled to follow the track (for more information on the track see above).
Then, fixing the trajectory for drone 1, the trajectory for drone 0 is optimized.
Next, vice versa, the trajectory for drone 1 is optimized while keeping the trajectory for drone 0 fixed.
This is done iteratively for a fixed number of iterations.
The hope is that eventually we arrive at a fixed-point, i. e. after optimizing for both drones we get the same
trajectories again.
COMPUTING THE BEST RESPONSE:
Since some of the constraints are non-convex quadratic constraint (namely the non-collision constraint),
the optimized trajectory is found in an iteratively fashion (sequential quadratic program, short SQP) by linearizing
the non-collision constraints around the current guess.
In case no feasible solution is found, the problem is relaxed by turning constraints into objectives.
PARAMETERS:
i_ego The index of the 'ego' drone
i_opp The index of the opponent drone
dt Sampling time for trajectories
blocking_term Coefficient on the "blocking" objective
n_steps Length of the trajectories
n_game_iters Number of game iterations (how often the best response is computed for drone i_0)
n_sqp_iters Number of SQP iterations (how often the constraints are linearized and the optimization is solved)
drone_paramstangents
r_coll Collision radius, see competition guidelines.
r_safe Safety radius, see competition guidelines
v_max Maximal velocity, determines how far waypoints can be apart
a_max Maximal acceleration, not used here.
"""
def __init__(self, params, drone_params, gate_poses):
self.dt = params.dt
self.n_steps = params.n
self.blocking = params.blocking
self.drone_params = drone_params
self.track = SplinedTrack(gate_poses)
# These are some parameters that could be tuned.
# They control how the safety penalty and the relaxed constraints are weighted.
self.nc_weight = 2.0
self.nc_relax_weight = 128.0
self.track_relax_weight = 128.0 # possibly should be the largest of the gains?
self.blocking_weight = 16.0 # increasing gain increases the aggressiveness in the blocking behavior
def init_trajectory(self, i_0, p_0):
"""Initialize Trajectory along the track tangent
Based on the start position p_0, return an initial guess for a trajectory
This is simply a line following the tangent of the track with maximal speed
:param i_0: The index of the drone to initialize the trajectory for.
:param p_0: The current position of the drone
:return: A trajectory of length self.n_steps
"""
v_ego = self.drone_params[i_0]["v_max"]
trajectory = np.zeros(shape=(self.n_steps, 3))
p = np.copy(p_0) # Copy state as it gets modified throughout the loop
for k in range(self.n_steps):
idx, c, t, n, width, height = self.track.track_frame_at(p)
p += self.dt * v_ego * t
p[2] = c[2] # fix trajectory height to center of track
trajectory[k, :] = p
return trajectory
def best_response(self, i_ego, state, trajectories):
"""Based on current trajectories, compute the best-response (BR) of player i_ego.
This is done by solving an optimization problem over the trajectory
p_ego[0], p_ego[1], ..., p_ego[N]
for drone i_ego maximizing the progress along the track while also trying to block the opponent. Here t is the tangent vector that points along the track and N is the horizon length. The resulting optimization is implemented as the following:
minimize -t^T p_ego[N]
subject to
- dynamical constraints
||p_ego[k+1] - p_ego[k]|| <= v_max*dt
- stay-within-track constraints
|n^T (p[k] - c)| <= width,
|v^T (p[k] - c)| <= height,
where v is the track vertical (t x n, where x here represents the cross product) and c is the track center
- non-collision constraints
||p_ego[k] - p_opp[k]|| >= r_coll * 2
The progress along the track is approximated by the progress along the tangent of the last point of the trajectory, i. e., maximize t^T p_ego[k]. The non-collision constraints are non convex and are linearized here. Instead of requiring the ego drone to stay outside a circle, the drone is constrained to be in a half-plane tangential to that circle. In addition to optimizing track progress, the penalty of violating the safety radius is accounted for.
If the blocking term defined in the trajectory parameters is non-zero, we add an additional term to the objective function that incentivizes the drones to slightly adjust their trajectories to block the opponent. Now the objective function is
-t^T p_ego[N] + sum_k( gamma^k p_rel[k]^T n n^T p_rel[k] )
where the sum here is over the full trajectory (1, ..., k, ..., N), gamma is the blocking coefficient (a term that is positive when the opponent is behind the ego agent and zero otherwise), and p_rel = p_ego - p_opp is the relative pose vector between the two drones. This is just a heuristic for "blocking" behavior that is only activated when the ego drone is in front of the opponent.
:param i_ego: The drone index of the ego drone.
:param state: The current state (positions) of the two drones.
:param trajectories: The current trajectories
:return: Optimized trajectory of player i_ego
"""
i_opp = (i_ego + 1) % 2
v_ego = self.drone_params[i_ego]["v_max"]
r_coll_ego = self.drone_params[i_ego]["r_coll"]
r_coll_opp = self.drone_params[i_opp]["r_coll"]
r_safe_ego = self.drone_params[i_ego]["r_safe"]
r_safe_opp = self.drone_params[i_opp]["r_safe"]
d_coll = r_coll_ego + r_coll_opp
d_safe = r_safe_ego + r_safe_opp
p = cp.Variable(shape=(self.n_steps, 3))
# === Dynamical Constraints ===
# ||p_0 - p[0]|| <= v*dt
init_dyn_constraint = cp.SOC(cp.Constant(v_ego * self.dt), cp.Constant(state[i_ego, :]) - p[0, :])
# ||p[k+1] - p[k]|| <= v*dt
dyn_constraints = [init_dyn_constraint] + [
cp.SOC(cp.Constant(v_ego * self.dt), p[k + 1, :] - p[k, :]) for k in range(self.n_steps - 1)]
# === Track Constraints ===
track_constraints = []
track_obj = cp.Constant(0)
track_objective_exp = 0.5 # exponentially decreasing weight
t = np.zeros((self.n_steps, 3))
n = np.zeros((self.n_steps, 3))
for k in range(self.n_steps):
# query track indices at ego position
idx, c, t[k, :], n[k, :], width, height = self.track.track_frame_at(trajectories[i_ego][k, :])
# hortizontal track height constraints
track_constraints.append(n[k, :].T @ p[k, :] - np.dot(n[k, :], c) <= width - r_coll_ego)
track_constraints.append(n[k, :].T @ p[k, :] - np.dot(n[k, :], c) >= -(width - r_coll_ego))
# vertical track height constraints
v = np.cross(t[k, :], n[k, :]) # the vertical direction component of the track
track_constraints.append(v.T @ p[k, :] - v.dot(c) <= height - r_coll_ego)
track_constraints.append(v.T @ p[k, :] - v.dot(c) >= -(height - r_coll_ego))
# track constraints objective
track_obj += (track_objective_exp ** k) * (
cp.pos(n[k, :].T @ p[k, :] - np.dot(n[k, :], c) - (width - r_coll_ego)) +
cp.pos(-(n[k, :].T @ p[k, :] - np.dot(n[k, :], c) + (width - r_coll_ego))))
# === Non-Collision Constraints ===
nc_constraints = []
nc_obj = cp.Constant(0)
nc_relax_obj = cp.Constant(0)
non_collision_objective_exp = 0.5 # exponentially decreasing weight
for k in range(self.n_steps):
p_opp = trajectories[i_opp][k, :]
p_ego = trajectories[i_ego][k, :]
# Compute beta, the normal direction vector pointing from the ego's drone position to the opponent's
beta = p_opp - p_ego
if np.linalg.norm(beta) >= 1e-6:
# Only normalize if norm is large enough
beta /= np.linalg.norm(beta)
# n.T * (p_opp - p_ego) >= d_coll
nc_constraints.append(beta.dot(p_opp) - beta.T @ p[k, :] >= d_coll)
# For normal non-collision objective use safety distance
nc_obj += (non_collision_objective_exp ** k) * cp.pos(d_safe - (beta.dot(p_opp) - beta.T @ p[k, :]))
# For relaxed non-collision objective use collision distance
nc_relax_obj += (non_collision_objective_exp ** k) * cp.pos(d_coll - (beta.dot(p_opp) - beta.T @ p[k, :]))
# === Blocking Heuristic Objective ===
blocking_obj = cp.Constant(0)
blocking_objective_exp = 0.5 # exponentially decreasing weight
leader_term = np.dot((trajectories[i_ego][0, :] - trajectories[i_opp][0, :]), t[0, :])
if ( self.blocking & (leader_term > 0.0) ):
for k in range(self.n_steps):
p_opp = trajectories[i_opp][k, :]
# scale factor for leading robot
p_rel = trajectories[i_ego][k, :] - p_opp
leader_term = np.dot(p_rel, t[k, :]);
gamma = 0.0
if (leader_term > 0):
gamma = 1.0/(leader_term * leader_term)/(k + 1);
else:
gamma = 0.0
# add blocking cost function
blocking_obj += gamma * blocking_objective_exp**k * cp.quad_form(p[k, :] - p_opp, np.outer(n[k, :], n[k, :]))
# === "Win the Race" Objective ===
# Take the tangent t at the last trajectory point
# This serves as an approximation to the total track progress
obj = -t[-1, :].T @ p[-1, :]
# create the problem in cxvpy and solve it
prob = cp.Problem(cp.Minimize(obj + self.nc_weight * nc_obj + self.blocking_weight * blocking_obj), dyn_constraints + track_constraints + nc_constraints)
# try to solve proposed problem
trajectory_result = np.array((self.n_steps, 3))
try:
prob.solve()
# relax track constraints if problem is infeasible
if np.isinf(prob.value):
print("WARN: relaxing track constraints")
# If the problem is not feasible, relax track constraints
# Assert it is indeed an infeasible problem and not unbounded (in which case value is -inf).
# (The dynamical constraints keep the problem bounded.)
assert prob.value >= 0.0
# Solve relaxed problem (track constraint -> track objective)
relaxed_prob = cp.Problem(cp.Minimize(obj + self.nc_weight * nc_obj + self.track_relax_weight * track_obj),
dyn_constraints + nc_constraints)
relaxed_prob.solve()
# relax non-collision constraints if problem is still infeasible
if np.isinf(relaxed_prob.value):
print("WARN: relaxing non collision constraints")
# If the problem is still infeasible, relax non-collision constraints
# Again, assert it is indeed an infeasible problem and not unbounded (in which case value is -inf).
# (The dynamical constraints keep the problem bounded.)
assert relaxed_prob.value >= 0.0
# Solve relaxed problem (non-collision constraint -> non-collision objective)
relaxed_prob = cp.Problem(cp.Minimize(obj + self.nc_weight * nc_obj + self.nc_relax_weight * nc_relax_obj),
dyn_constraints + track_constraints)
relaxed_prob.solve()
assert not np.isinf(relaxed_prob.value)
trajectory_result = p.value
except: # if cvxpy fails, just return the initialized trajectory to do something
print("WARN: cvxpy failre: resorting to initial trajectory (no collision constraints!)")
trajectory_result = trajectories[i_ego]
return trajectory_result
def iterative_br(self, i_ego, state, n_game_iterations=2, n_sqp_iterations=3):
trajectories = [
self.init_trajectory(i, state[i, :]) for i in [0, 1]
]
t0 = time.time()
for i_game in range(n_game_iterations - 1):
for i in [i_ego, (i_ego + 1) % 2]:
for i_sqp in range(n_sqp_iterations - 1):
trajectories[i] = self.best_response(i, state, trajectories)
# one last time for i_ego
for i_sqp in range(n_sqp_iterations):
trajectories[i_ego] = self.best_response(i_ego, state, trajectories)
t1 = time.time()
print('Total IBR solution time: ', t1 - t0)
return trajectories[i_ego]
def truncate(self, p_i, trajectory):
"""
Truncates the trajectory at time k, so that the next point is 'ahead' of p_i.
A point p is ahead of p_i, if p-p_i projected onto the track tangent is positive
:param p_i: The position of the drone
:param trajectory: The trajectory to be truncated
:return: k, the index of the first point ahead of p_i
"""
_, _, t, _, _, _ = self.track.track_frame_at(p_i)
truncate_distance = 0.01 # could be a parameter based on max velocity and computation time
for k in range(self.n_steps):
if t.dot(trajectory[k, :] - p_i) > truncate_distance: # truncate if next waypoint is closer than truncate_distance meters in front of robot
return k, t
return self.n_steps, t